Upload README.md with huggingface_hub

README.md

@@ -14,11 +14,15 @@ facts the named person states about themselves, one short sentence per line,
 or NONE. sourcedrecall runs them inside its own process on onnxruntime-genai,
 on the CPU, with no server.
 
-| file | base | download | per conversation (4 CPU threads) | peak memory |
+| file | base | download | a chat session (4 CPU threads) | peak memory |
 |---|---|---|---|---|
 | `sourcedrecall-notes-en-0.6b.tar.gz` (small) | Qwen3-0.6B | 0.51 GB | 2-3 s | 1.9 GB |
 | `sourcedrecall-notes-en-1.7b.tar.gz` (standard) | Qwen3-1.7B | 1.3 GB | about 4.5 s | 3.3 GB |
 
+A long coding session (eight parts of 12,000 characters) takes about 100 s
+and up to 3.6 GB with the small model; sourcedrecall writes notes in the
+background after the session ends.
+
 They are optional. sourcedrecall stores and finds the user's own words
 without them; with one, short notes are added, each kept only if its words
 are in what the user wrote, and shown labelled as written by a model.
@@ -42,11 +46,11 @@ are in what the user wrote, and shown labelled as written by a model.
 
 ## Measured (local Qwen3-14B reader and judge for every system)
 
-LoCoMo, conversations 2-9 (1,307 questions):
+LoCoMo, conversations 2-9 (1,307 questions; used to choose between models, so not held out):
 
 | | overall | multi-hop | temporal | single-hop |
 |---|---|---|---|---|
-| sourcedrecall, no notes | 64.7% | 53.6% | 56.2% | 74.8% |
+| sourcedrecall, no notes | 64.7% | 52.3% | 53.5% | 75.5% |
 | with the small model's notes | 68.2% | 55.2% | 58.9% | 79.0% |
 | with the standard model's notes | 70.3% | 59.0% | 57.4% | 81.6% |
 | with Qwen3-14B's notes (a server) | 73.1% | 62.3% | 57.4% | 85.4% |
@@ -65,8 +69,13 @@ published LoCoMo results, which use GPT-4-class models.
 ## Limits
 
 English only. They sometimes give the user something the other person said
-("has kids" from the other person's children). Neither closes
-sourcedrecall's multi-hop gap to Mem0.
+("has kids" from the other person's children), or keep a note whose words
+the user wrote but whose meaning differs (a negation, a number). The small
+model often writes nothing for short sessions; standard does better there.
+Neither closes sourcedrecall's multi-hop gap to Mem0. The held-out false
+memory sets were run many times during development; v4 is no longer clean.
+Known issues and fixes in progress: docs/REVIEW_2026-10-09.md in the
+sourcedrecall repository.
 
 ## Attribution